Blast radius

Blast radius is how far the damage spreads if an agent action goes wrong: how many people, files, or systems a single mistake would touch. Sorting actions by blast radius decides which an agent may run alone and which need review.

Blast radius is how far the damage spreads if an agent action goes wrong: how many people, files, or systems a single mistake would touch. A low-blast-radius action stays contained, like renaming one of your own draft notes; a high-blast-radius action reaches far, like emailing two hundred customers or deleting a shared folder. Say your agent tidies a to-do list. Marking one task done affects only you and is easy to undo, so the blast radius is small. Sending a calendar invite to a whole team reaches many people and is awkward to unwind, so the blast radius is large. You tell the agent to act freely on the small-radius steps and to pause for your approval on the large-radius ones.

Builder example

Estimating blast radius before you grant autonomy keeps a small bug from becoming a public incident. An agent that cleans up duplicate contacts can run unattended, because a wrong merge touches one record you can fix later. An agent that sends invoices should stop for review, because one bad run reaches every customer and is hard to recall. Tell the agent which categories of action are low-radius and safe to complete on its own, and which are high-radius and must wait behind your approval.

Common confusion: Blast radius measures how widely a failure would spread, while how often it fails is a separate question. A rare action with a huge reach, like wiping a shared database, still deserves an approval gate. What sets the gate is the size of the worst outcome, not how likely it is on any single run.